Although recent experiments have shown that visual search can be based on features other than simple 2D image patterns (e.g., surface convexity), the assumption has remained that the space in which these features are embedded is relatively simple. In particular, it is often assumed that this “embedding space” is retinotopic, or is a spatiotopic one that only factors out retinal shifts. To determine the nature of this representation, a “shaker paradigm” was developed, in which an initial display alternated with a geometrically transformed version of itself several times a second. Performance for this was compared against that for the same task carried out on a static display. To the degree that a task is based upon a representation invariant to the transformation applied, performance will be unaffected; otherwise, it will require switching between two separate representations. Search was for a T-shaped target among L-shaped distractors. To test invariance to rotation, a first set of experiments used displays that rotated back and forth by a fixed angle. A range of angles (7-60 deg) was tested, as was a range of alternation times (250-750 ms). Slopes and intercepts were unaffected by rotations of up to 20 degrees, with performance deteriorating sharply beyond that limit. A second set of experiments tested invariance to size scaling. A range of scales (4:3-4:1) was tested, as was a range of alternation times (250-750 ms). Slopes and intercepts were unaffected by scale changes of up to 2:1, with performance deteriorating for scale changes of 3:1 or greater. These results show that visual search is based on an embedding space that is not highly invariant to rotation. However, it is largely invariant to scaling, indicating that it is more abstract than a retinotopic or simple spatiotopic representation.

Support provided by the Natural Sciences and Engineering Research Council of Canada and Nissan Motor Co., Ltd.